Feature matching between images in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - Feature matching between images

Which metric matters for feature matching and WHY

In feature matching between images, the key metrics are matching accuracy and matching precision. These tell us how many matched points are correct versus incorrect. Since feature matching finds pairs of points that represent the same real-world spot, we want to measure how many matches are true (correct) and how many are false (wrong). High precision means most matches are correct, which is important to avoid wrong matches that confuse later steps like 3D reconstruction or stitching.

Confusion matrix for feature matching

      | Predicted Match    | Predicted No Match |
      |--------------------|--------------------|
      | True Positive (TP)  | False Negative (FN) |
      | False Positive (FP) | True Negative (TN)  |

      TP: Correctly matched points
      FP: Incorrectly matched points
      FN: Missed correct matches
      TN: Correctly rejected non-matches
    

For example, if we have 100 true matching points, and our algorithm finds 90 matches, 80 of which are correct (TP=80), 10 wrong (FP=10), and misses 20 (FN=20), then:

• Precision = 80 / (80 + 10) = 0.89
• Recall = 80 / (80 + 20) = 0.80

Precision vs Recall tradeoff with examples

Feature matching often balances precision and recall:

• High precision, low recall: Matches are mostly correct but many true matches are missed. Useful when wrong matches cause big problems, like in 3D modeling where errors ruin the model.
• High recall, low precision: Most true matches are found but many wrong matches appear. This might be okay for rough alignment but can cause errors downstream.

For example, in panorama stitching, high precision avoids visible ghosting from wrong matches. In object recognition, high recall ensures the object is detected even if some matches are noisy.

What good vs bad metric values look like

Good feature matching results:

• Precision above 0.85 means most matches are correct.
• Recall above 0.75 means most true matches are found.
• Balanced precision and recall around 0.8 or higher is ideal.

Bad results:

• Precision below 0.5 means many wrong matches, causing errors.
• Recall below 0.4 means many true matches are missed, losing information.
• Very high recall but very low precision means noisy matches.

Common pitfalls in feature matching metrics

• Ignoring false matches: Only counting total matches without checking correctness can be misleading.
• Data leakage: Using the same images for tuning and testing inflates metrics.
• Overfitting: Matching tuned to specific image pairs may fail on new images.
• Accuracy paradox: High overall match count but low precision means many wrong matches.

Self-check question

Your feature matching model finds 95% of true matches (recall = 0.95) but only 40% of its matches are correct (precision = 0.40). Is this good for production? Why or why not?

Answer: No, this is not good. Although the model finds most true matches, it also produces many wrong matches (low precision). Wrong matches can cause errors in later steps like stitching or 3D reconstruction. You want to improve precision to reduce false matches.

Key Result

Precision and recall are key to evaluate feature matching; high precision avoids wrong matches, high recall finds most true matches.

Practice

1. What is the main purpose of feature matching between two images?

easy

A. To find similar points or patterns between the images

B. To change the colors of the images

C. To increase the image resolution

D. To crop the images automatically

Solution

1. Step 1: Understand feature matching concept

   Feature matching is used to find points in two images that look alike, such as corners or edges.

2. Step 2: Identify the main goal

   The goal is to find these similar points to compare or align images, not to change colors or resolution.

3. Final Answer:

   To find similar points or patterns between the images -> Option A

4. Quick Check:

   Feature matching = find similar points [OK]

Hint: Feature matching finds points that look alike in two images [OK]

Common Mistakes:

• Confusing feature matching with image editing
• Thinking it changes image size or colors
• Mixing feature matching with image cropping

2. Which of the following is the correct way to detect keypoints using ORB in OpenCV (Python)?

easy

A. orb = cv2.ORB_create(); keypoints = orb.getKeypoints(image)

B. orb = cv2.ORB(); keypoints = orb.find(image)

C. orb = cv2.ORB_create(); keypoints = orb.detect(image, None)

D. orb = cv2.ORB_create(); keypoints = orb.findKeypoints(image)

Solution

1. Step 1: Recall ORB keypoint detection syntax

   In OpenCV, ORB keypoints are detected using orb = cv2.ORB_create() and orb.detect(image, None).

2. Step 2: Check each option

   orb = cv2.ORB_create(); keypoints = orb.detect(image, None) matches the correct syntax; others use incorrect method names or constructors.

3. Final Answer:

   orb = cv2.ORB_create(); keypoints = orb.detect(image, None) -> Option C

4. Quick Check:

   Correct ORB syntax = orb = cv2.ORB_create(); keypoints = orb.detect(image, None) [OK]

Hint: Use ORB_create() and detect() to find keypoints [OK]

Common Mistakes:

• Using wrong method names like findKeypoints
• Calling ORB() instead of ORB_create()
• Passing wrong arguments to detect()

3. Given the following code snippet, what will be the output length of good_matches?

import cv2
orb = cv2.ORB_create()
kp1, des1 = orb.detectAndCompute(img1, None)
kp2, des2 = orb.detectAndCompute(img2, None)
matcher = cv2.BFMatcher(cv2.NORM_HAMMING)
matches = matcher.knnMatch(des1, des2, k=2)
good_matches = []
for m, n in matches:
    if m.distance < 0.75 * n.distance:
        good_matches.append(m)
print(len(good_matches))

medium

A. Total number of keypoints in img2

B. Number of matches passing the ratio test

C. Total number of keypoints in img1

D. Total number of all matches found

Solution

1. Step 1: Understand knnMatch and ratio test

   knnMatch finds the two best matches for each descriptor. The ratio test keeps matches where the best is significantly better than the second best.

2. Step 2: Analyze the code logic

   The loop filters matches by distance ratio, so good_matches contains only those passing the test, not all matches or keypoints.

3. Final Answer:

   Number of matches passing the ratio test -> Option B

4. Quick Check:

   good_matches length = matches passing ratio test [OK]

Hint: Ratio test filters matches; good_matches count = filtered matches [OK]

Common Mistakes:

• Confusing matches with keypoints count
• Thinking good_matches includes all matches
• Ignoring the ratio test condition

4. Identify the error in this feature matching code snippet:

import cv2
orb = cv2.ORB_create()
kp1, des1 = orb.detectAndCompute(img1, None)
kp2, des2 = orb.detectAndCompute(img2, None)
matcher = cv2.BFMatcher(cv2.NORM_L2)
matches = matcher.match(des1, des2)
print(len(matches))

medium

A. Using cv2.NORM_L2 with ORB descriptors is incorrect

B. Missing call to detect() before detectAndCompute()

C. BFMatcher should be replaced with FlannBasedMatcher

D. match() requires k parameter for ORB descriptors

Solution

1. Step 1: Check descriptor type and matcher norm

   ORB descriptors are binary, so BFMatcher should use cv2.NORM_HAMMING, not NORM_L2.

2. Step 2: Identify the error

   Using NORM_L2 causes incorrect distance calculation and poor matching for ORB.

3. Final Answer:

   Using cv2.NORM_L2 with ORB descriptors is incorrect -> Option A

4. Quick Check:

   ORB needs NORM_HAMMING, not NORM_L2 [OK]

Hint: Use NORM_HAMMING with ORB descriptors [OK]

Common Mistakes:

• Using wrong norm type for binary descriptors
• Thinking detect() is needed before detectAndCompute()
• Confusing BFMatcher with FlannBasedMatcher

5. You want to match features between two images taken from different angles. Which approach improves matching accuracy the most?

hard

A. Use ORB detector without any filtering on matches

B. Resize images to very small size before matching

C. Use random keypoints and brute force matching

D. Use SIFT detector and apply Lowe's ratio test on matches

Solution

1. Step 1: Consider feature detector choice

   SIFT is robust to scale and rotation changes, better for different angles than ORB or random points.

2. Step 2: Apply filtering for accuracy

   Lowe's ratio test filters out weak matches, improving accuracy significantly.

3. Step 3: Evaluate other options

   Using ORB without filtering or random points reduces accuracy; resizing too small loses details.

4. Final Answer:

   Use SIFT detector and apply Lowe's ratio test on matches -> Option D

5. Quick Check:

   SIFT + ratio test = best accuracy [OK]

Hint: SIFT + Lowe's ratio test improves matching accuracy [OK]

Common Mistakes:

• Skipping ratio test filtering
• Using random or weak keypoints
• Reducing image size too much